AUTHOR=Yang Jian , Zhang Tianyu , Zhang Junping , Lin Xun , Wang Hailong , Feng Tao TITLE=A ConvLSTM nearshore water level prediction model with integrated attention mechanism JOURNAL=Frontiers in Marine Science VOLUME=11 YEAR=2024 URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2024.1470320 DOI=10.3389/fmars.2024.1470320 ISSN=2296-7745 ABSTRACT=

Nearshore water-level prediction has a substantial impact on the daily lives of coastal residents, fishing operations, and disaster prevention and mitigation. Compared to the limitations and high costs of traditional empirical forecasts and numerical models for nearshore water-level prediction, data-driven artificial intelligence methods can more efficiently predict water levels. Attention mechanisms have recently shown great potential in natural language processing and video prediction. Convolutional long short-term memory(ConvLSTM) combines the advantages of convolutional neural networks (CNN) and long short-term Memory (LSTM), enabling more effective data feature extraction. Therefore, this study proposes a ConvLSTM nearshore water level prediction model that incorporates an attention mechanism. The ConvLSTM model extracts multiscale information from historical water levels, and the attention mechanism enhances the importance of key features, thereby improving the prediction accuracy and timeliness. The model structure was determined through experiments and relevant previous studies using five years of water level data from the Zhuhai Tide Station and the surrounding wind speed and rainfall data for training and evaluation. The results indicate that this model outperforms the four other baseline models of PCCs, RMSE, and MAE, effectively predicting future water levels at nearshore stations up to 48 h in advance. Compared to the ConvLSTM model, the model with the attention mechanism showed an average improvement of approximately 10% on the test set, with a greater error reduction in short-term forecasts than that in long-term forecasts. During Typhoon Higos, the model reduced the MAE of the best-performing baseline model by approximately 3.2 and 2.4 cm for the 6- and 24-hour forecasts, respectively, decreasing forecast errors by approximately 18% and 11%, effectively enhancing the ability of the model to forecast storm surges. This method provides a new approach for forecasting nearshore tides and storm surges.